Reliable and fast hand-offs in low-power wireless networks

Reliable and fast hand-offs in low-power wireless networks

Technical Report

*CISTER Research Center

CISTER-TR-140501

26, Sep, 2014

Hossein Fotouhi*

Mário Alves*

Marco Zuniga

Anis Koubâa*

Technical Report CISTER-TR-140501 Reliable and fast hand-offs in low-power wireless networks

© CISTER Research Center www.cister.isep.ipp.pt 1

Reliable and fast hand-offs in low-power wireless networks Hossein Fotouhi*, Mário Alves*, Marco Zuniga, Anis Koubâa*

*CISTER Research Center

Polytechnic Institute of Porto (ISEP-IPP)

Rua Dr. António Bernardino de Almeida, 431

4200-072 Porto

Portugal

Tel.: +351.22.8340509, Fax: +351.22.8340509

E-mail: [email protected], [email protected], [email protected]

http://www.cister.isep.ipp.pt

Abstract Hand-off (or hand-over), the process where mobile nodes select the best access point available to transfer data, has been well studied in wireless networks. The performance of a hand-off process depends on the specific characteristics of the wireless link. In the case of low-power wireless networks, hand-off decisions must be carefully taken by considering the unique properties of inexpensive low-power radios. This article addresses the design, implementation and evaluation of smart-HOP, a hand-off mechanism tailored for low-power wireless networks. This work has three main contributions. First, it formulates the hard hand-off process for low-power networks (such as typical wireless sensor networks -WSNs) with a probabilistic model, to investigate the impact of the most relevant channel parameters through an analytical approach. Second, it confirms the probabilistic model through simulation and further elaborates on the impact of several hand-off parameters. Third, it fine-tunes the most relevant hand-off parameters via an extended set of experiments, in a more realistic experimental scenario. The evaluation shows that smart-HOP performs well in the transitional region while achieving more than 98% relative delivery ratio and hand-off delays in the order of a tenth of a second.

IEEE TRANSACTIONS ON MOBILE COMPUTING 1

Reliable and fast hand-offs in low-powerwireless networks

Hossein Fotouhi, Mario Alves, Marco Zuniga, and Anis Koubaa,

Abstract—Hand-off (or hand-over), the process where mobile nodes select the best access point available to transfer data, hasbeen well studied in wireless networks. The performance of a hand-off process depends on the specific characteristics of thewireless links. In the case of low-power wireless networks, hand-off decisions must be carefully taken by considering the uniqueproperties of inexpensive low-power radios. This article addresses the design, implementation and evaluation of smart-HOP, ahand-off mechanism tailored for low-power wireless networks. This work has three main contributions. First, it formulates thehard hand-off process for low-power networks (such as typical wireless sensor networks - WSNs) with a probabilistic model,to investigate the impact of the most relevant channel parameters through an analytical approach. Second, it confirms theprobabilistic model through simulation and further elaborates on the impact of several hand-off parameters. Third, it fine-tunesthe most relevant hand-off parameters via an extended set of experiments, in a realistic experimental scenario. The evaluationshows that smart-HOP performs well in the transitional region while achieving more than 98% relative delivery ratio and hand-offdelays in the order of a few tens of a milliseconds.

Index Terms—Mobility, Low-power links, Wireless sensor networks, Link characteristics, hand-off, hand-over.

F

1 INTRODUCTION

WIRELESS technologies have been key enablerfor an expanding range of mobile applica-

tions, building not only on smart phones and tablets,but also on wearable sensors, industrial machinery,health-monitoring instruments and robotics [1]. Thesedevices play an instrumental role in many new appli-cation domains that push wireless networks to dra-matically improve quality-of-service properties suchas throughput, timeliness, reliability, security, privacy,usability, and efficiency [1]–[3]. These QoS require-ments must be guaranteed between mobile nodesand also between mobile nodes and fixed networkinfrastructures.

A recent NSF report [1] presents an exemplaryscenario capturing this situation. In the future, in-body sensors collecting “aggregated data from the entirepopulation can predict outbreaks of epidemics even beforethey occur” and it further states that “such applicationsrequire in-body sensors that not only have robust wirelessconnectivity, but also are highly energy-efficient”. Wecan easily foresee a hospital covered by a wirelesssensor network (WSN) infrastructure (used for one ormore purposes). Patients use body sensor networksto monitor relevant/vital signs —they are monitoredand tracked either when standing still or moving(walking, wheel chair or in bed). Doctors and medicalstaff also use body sensor networks, which can beused to measure their stress/anxiety levels and also

• Hossein Fotouhi and Mario Alves are with the CISTER/INESC-TEC,ISEP, Polytechnic Institute of Porto, Portugal.E-mails: (mohfg,mjf)@isep.ipp.pt

• Marco Zuniga is with the Embedded Software Group, Delft Universityof Technology, The Netherlands.E-mail: [email protected]

• Anis Koubaa is with the CISTER/INESC-TEC, ISEP, PolytechnicInstitute of Porto, Portugal and COINS Research Group, Prince SultanUniversity, Saudi Arabia.E-mail: [email protected]

to track and warn them about emergency situations.Only if these wearable body sensor nodes are able tocommunicate reliably and in real-time, this will effec-tively transform medical services in the near future.

In industrial environments such as factory automa-tion and process control, it is essential to monitorthe actual state of components and machines in acontinuous manner. The Factories of the Future 2020roadmap [2] forecasts “the need for advanced machineinteraction with humans through ubiquity of mobile devicesto receive relevant production information”. Such type ofsystems is also expected to detect potentially dan-gerous conditions in real-time and launch necessarycountermeasures to prevent their impact on workers’health and safety.

The Cooperating Objects roadmap [3] envisions andoutlooks several application domains requiring thecooperation between mobile robots instrumented withsensing/actuation capabilities with fixed wireless sen-sor nodes, such as for search & rescue, environmentexploration and surveillance applications. A largenumber of small (and inexpensive) robots can cooper-ate to tackle a large problem. These swarms of robotspose important challenges to robot designers as theircooperative behavior is not as simple to program as asingle robot: many algorithms are distributed and relyheavily on communication between the participatingmembers of the swarm and also with a fixed wirelessinfrastructure. Typically, this communication is time-critical, meaning it has to be completed within a timedeadline to be effective.

Many recent research projects (e.g. [4]–[7]) and re-search works (e.g. [8]–[12] have considered networkarchitectures that require real-time (or at least con-tinuous) data collection from mobile nodes throughlow-power wireless interfaces to fixed network in-frastructures. In oil refineries, workers are exposedto hazardous environments in highly critical areas,so collecting the workers’ vital signs during theirdaily activity enables to quickly detect abnormal sit-


uations [8]. In clinical monitoring, patients have em-bedded sensing devices that report real-time streamsof information through a fixed infrastructure [9], [10].Mobile robots are also used to assist fixed sensornetwork deployments in wildlife monitoring to detectand extinguish fire [11], [12].

The communication between mobile nodes and afixed infrastructure has been extensively studied inCellular and WiFi networks, and it has been ad-dressed through the use of hand-off mechanisms.However, these methods cannot be readily applied tolow-power wireless networks [13]. First, Cellular andWiFi networks have more sophisticated radios withmore energy resources. This means that their wirelesslinks are much longer and more reliable than thoseprovided by low-power low-cost radios, and hencethe thresholds and parameters associated to hand-offmechanisms need to be tuned accordingly. Second,base stations in cellular networks build on fixed wiredinfrastructures with strong processing and communi-cation capabilities, which is usually not applicable inlow-power networks. Third, mobile nodes in Cellularand WiFi networks are usually in the coverage rangeof several strong radios while the unreliable links oflow-power wireless networks have little overlap.

In this paper we address the design, implementa-tion and evaluation of smart-HOP, a hand-off mecha-nism that considers the specific features of low-powerlinks to enable fast, reliable and efficient hand-offs1.We enhance our preliminary work published at [13]with the following new contributions:

1) We formulate a hard hand-off process for low-power networks with a probabilistic model tostudy the impact of relevant channel parameters.

2) We design a simulation model to confirm theprobabilistic analysis and also to analyze theimpact of relevant network parameters on theoverall performance.

3) We further fine-tune the hand-off parametersthrough an extensive set of experiments in arealistic environment with a person holding themobile node.

Organization. In Section 2, we explain the mainlimitations of low-power networks and overviewsome hand-off approaches for low-power wirelessnetworks. In the remainder of the paper, the termslow-power wireless networks and wireless sensor net-works (WSNs) are used interchangeably. In Section 3,we describe the smart-HOP mechanism and its mainparameters, and illustrate some experimental resultsobtained in a controlled environment. The analyticaland simulation models together with an extensivestudy of the impact of channel parameters are pre-sented in Section 4. In Section 5, we provide the bestparameter tuning based on an extensive experimentalanalysis in a realistic environment. Related work isoutlined in Section 6. Finally, we conclude the paperand discuss our most relevant findings in Section 7.

1The extended version of this paper is available online [14].

2 PROBLEM STATEMENTThis section elaborates on the need to calibrate hand-offs according to the particular characteristics of low-power wireless networks and on the parameters thatshould be taken into account when designing a hand-off mechanism.

A naive solution to support mobility in WSNs isfor mobile nodes (MNs) to broadcast messages to allaccess points (APs) in their vicinity. These broadcastapproach, while simple, has a major limitation: broad-casts lead to redundant information at neighboringAPs (since more than one AP may receive the samepacket). This implies that the fixed infrastructure hasto either waste resources in forwarding the sameinformation to the end point, or to use a complexscheme, such as data fusion, to eliminate duplicatedpackets locally. A more efficient solution is for mobilenodes to select a single AP to transmit data at anygiven time. This alternative requires nodes to per-form hand-offs between neighboring APs. Contrarilyto more powerful wireless systems, such as cellu-lar networks, which have typically advanced spreadspectrum radios and high energy resources, WSNshave severely constrained resources. Hence, we needa better understanding of the hand-off process in low-power wireless networks.

Limitations of low-power links. Low-power linkshave two characteristics that affect the hand-off pro-cess: short coverage and high variability [15], [16].Several empirical studies revealed the existence ofthree distinct reception regions in a wireless link;connected, transitional, and disconnected [17]. Thetransitional region is often quite significant in size,and is generally characterized by high variance inreception rates and asymmetric connectivity. In WSNapplications, most of the links (more than 50% [15])are in the transitional region.

Studies show that WSN links have high unreliabil-ity in dense deployments [18], [19]. The high vari-ability of links has a direct impact on the stabilityof hand-offs. When not designed properly, hand-offmechanisms may degrade the network performancedue to the ping-pong effect, which consists in mobilenodes having consecutive and redundant hand-offsbetween two APs due to sudden fluctuations of theirlink qualities. This usually happens when a mobilenode moves in the vicinity of two APs. Hence, to effec-tively cope with link instability, a hand-off mechanismshould calibrate the appropriate thresholds, takinginto account the variance of the wireless links.

The transitional region for wireless nodes using theCC2420 radio transceiver encompasses the approxi-mate range [-92 dBm, -80 dBm]. Intuition may dictatethat the hand-off should be performed within theconnected region as it indicates more reliable links. Inpractice, a hand-off should start when the link withthe current (serving) AP drops below a given value(Tl) and should stop when it finds a new AP with therequired link quality (above Th).

Figure 1(a) depicts an example of inefficient hand-off and illustrates the negative impact of this conser-vative approach. In this scenario, the lower thresholdis set to -85 dBm, and the upper threshold is set


0 1 2 3 4−95

−90

−85

−80

−75

−70

−65

−60

RSS

I (dB

m)

Time (s)

TH (high)

TH (low)

Handoff(ping−pong

effect)

0 1 2 3 4−95

−90

−85

−80

−75

−70

−65

−60

RSS

I (dB

m)

Time (s)

Handoff

TH (high)

TH (low)

(a) (b)

Fig. 1. (a) an example of inefficient hand-off withnarrow hysteresis margin (1 dBm), Tl = �86 dBmand Th = �85 dBm, resulting in three consecutivehand-offs (ping-pong effect). (b) an example of anefficient hand-off with wide hysteresis margin (5 dBm),Tl = �90 dBm and Th = �85 dBm, resulting in a singlehand-off [13].

to 1 dBm higher. This particular choice of param-eters results in three undesirable consecutive hand-offs between the two contiguous APs (three shadowedvertical bars), which we refer to as ping-pong effect andthat results in a long network inaccessibility time (700ms). Increasing the threshold margin to 5 dBm (asillustrated in Figure 1(b)) eliminates the ping-pongeffect (one shadowed vertical bar only) and hencereduces the hand-off delay to approximately 200 ms.This simple example shows that studying the low-power link characteristics is paramount for obtainingefficient hand-offs.

3 BASICS ON THE SMART-HOPIn this section, first we explain the data communi-cation between the MN and a fixed infrastructure ofAPs, according to the smart-HOP procedure. Then,we highlight the importance of three parameters:window size, hysteresis threshold and stability monitoring.Afterward, we compare the cost of communication insmart-HOP with conventional hand-off approaches.Evaluation based on experiments in a controlled en-vironment (model train moving in a square track) isthen discussed.

3.1 The smart-HOP algorithmThe smart-HOP algorithm has been proposed in [13].The algorithm has two main phases: (i) Data Transmis-sion Phase and (ii) Discovery Phase. A timeline of thealgorithm is depicted in Figure 2.

For the sake of clarity, let us assume that a node is inthe Data Transmission Phase2. In this phase, the mobilenode is assumed to have a reliable link with an AP,defined as serving AP in Figure 2. The mobile nodemonitors the link quality by receiving reply packetsfrom the serving AP. Upon receiving n data packetsin a given window, the serving AP replies with theaverage RSSI or SNR of the n packets. If no packetsare received, the AP takes no action. This may lead

2smart-HOP has a simple initialization phase that is similar tothe Discovery Phase.

to disconnections, which are solved through the useof a time-out mechanism. It is important to noticethat smart-HOP filters out asymmetric links implicitlyby using reply packets at the Data Transmission andDiscovery Phases. If a neighboring AP does not haveactive links in both directions, that AP is simply notpart of the process. The smart-HOP process relies onthree main parameters, follows.

Parameter 1: window size (ws). It represents thenumber of packets required to estimate the link qual-ity over a specific time interval3. A small ws (highsampling frequency) provides detailed informationabout the link but increases the processing of replypackets, which leads to higher energy consumptionand lower delivery rates. The packet delivery reducesas the MN opts for performing some unnecessaryhand-offs. The hand-off is triggered by detecting lowquality links, resulting from the decrease of the signalstrength. On the other hand, a large ws (low samplingfrequency) provides only coarse grained informationabout the link and decreases the responsiveness ofthe system, which is not suitable for mobile networkswith dynamic link changes.

Parameter 2: hysteresis margin (HM). In WSNs, theselection of thresholds and hysteresis margins is dic-tated by the characteristics of the transitional regionand the variability of the wireless link. The thresholdsshould be selected according to the boundaries of thetransitional region. The transitional region is oftenquite significant in size and hence a large numberof links in the network (higher than 50%) are unre-liable [20], [21]. Therefore, wireless nodes are likelyto spend most of the time in the transitional region.

A tight estimation of the threshold level withinthe transitional region is obtained from experimentalanalysis. If the Tl threshold is too high, the nodecould perform unnecessary hand-offs (by being tooselective). If the threshold is too low, the node mayuse unreliable links. The hysteresis margin plays a cen-tral role in coping with the variability of low-powerwireless links. If the hysteresis margin is too narrow,the mobile node may end up performing unnecessaryand frequent hand-offs between two APs (ping-pongeffect), as illustrated in Figure 1. If the hysteresis marginis too large, hand-offs may take too long, which endsup increasing the network inaccessibility times, andthus decreasing the delivery rate.

Parameter 3: stability monitoring (m). Due to thehigh variability of wireless links, the mobile node maydetect an AP that is momentarily above Th, but thelink quality may decrease shortly after handing-off tothat AP. In order to avoid this, it is important to assessthe stability of the candidate AP. After detecting anAP with the link quality above Th, the MN sends mfurther bursts of beacons to check the stability of thatAP. The burst of beacons stands for the ws requestbeacons followed by the reply packets received from

3In the extended experiments, the beacons’ interval increasedfrom 5 ms (in the preliminary experiments) to 10 ms. The longerperiod in transmitting beacons increased the chance of beaconreception at the APs. In the Data Transmission Phase, after sendinga burst of beacons, the MN waits for 10 ms to receive the replyfrom the serving parent. In the Discovery Phase, the waiting time isincreased to 100 ms in order to get replies from all neighbor APs.


MN

serving AP

Data TX

. . .Reply

RSSI � TH low

. . .

Data TXReply

RSSI<TH low

All APs

. . .

n Beacons

. . .

APs Reply

TDMA slots...

RSSI � TH high

Select best AP andgo to Data TX Phase

Data Transmission Phase

Discovery Phase

Fig. 2. Timing diagram of the smart-HOP mechanism [13].

the neighboring APs. As can be easily inferred, the sta-bility monitoring and the hysteresis margin parametersare tightly coupled . A wide hysteresis margin requiresa lower m, and vice-versa. In the experimental eval-uation (Section 5), we will show that an appropriatetuning of the hysteresis margin will lead to m = 1.

3.2 Why smart-HOP for WSNs?Hand-offs are used in all mobile wireless networks.This simple concept of switching from one AP toanother requires very careful design considerations,so that the application requirements and system lim-itations are respected. In this subsection, first weoutline the main features of hand-off processes inCellular and WiFi networks, and show that a newapproach is required for low-power wireless networks(smart-HOP), comparing the communication cost ofthese hand-off approaches.

In Cellular networks, the base stations have highenergy, processing and communication resources. Allthe APs are connected through a stable wired back-bone, which is responsible for making hand-off de-cisions. All mobile nodes periodically broadcast bea-cons along with their data packets. At the same time,the base stations communicate with each other andassess the location and the link quality of all mobilenodes. By detecting a low quality link, the base sta-tions decide for the next servicing base station (for theMN).

In WiFi networks, power and bandwidth are morelimited than in cellular networks. Thus, performinga centralized decision at the base stations (similarto cellular networks) is not efficient. In these net-works, a distributed hand-off decision is performedat the MNs. All APs periodically broadcast beacons invarious available channels with a precise timing (toeliminate overlapping). The MN periodically broad-casts request packets in all channels to get immediatereplies (beacons) from neighbor APs. During the DataTransmission Phase, the MN gets periodic beacons fromthe serving AP. By detecting a low quality link withthe serving AP and high quality link with one ofthe neighbors, the MN decides to trigger a hand-offprocess.

smart-HOP can reduce the communication over-head. Applying the aforementioned techniques inWSNs requires a lot of beaconing, which in turnincreases the network overhead, collisions and en-ergy consumption. In low-power low-cost wirelessnetworks with a poor backbone of APs, a centralized

approach is not feasible. On the other hand, a periodicbeaconing of APs in a single radio network leads topacket collisions. smart-HOP is a distributed hand-offmechanism where MNs are responsible for broadcast-ing beacons after detecting a low quality link.

Let us assume a simple terminology to depictthe communication overhead of smart-HOP. Denot-ing Ctx, Crx, Cb, and nAP as the transmission cost,reception cost, beaconing cost and average number ofAPs available4. The communication overhead of allwireless networks is formulated as follows. (i) WSNswith smart-HOP is (Ctx + Crx)(1 + 1

ws ), (ii) WSNswith broadcast approach is Ctx+nAP ⇥Crx, (iii) WiFinetworks is (Ctx + Crx) + (nAP ⇥ Cb + nAP ⇥ Crx),and (iv) cellular networks is (Ctx+nAP ⇥Crx)+(Cb+nAP ⇥ Crx). It is important to note that in estimatingthe costs, we considered the general concept of hand-off that is common in most of the literature. Simplemanipulations lead to the following conditions5.

1) Csmart-HOP > CBroadcast

if (ws⇥ nAP � ws� 1)Crx < Ctx

2) Csmart-HOP > Ccellular

if ws⇥ Cb + Crx(2nAP ⇥ ws� ws� 1) < Ctx

3) Csmart-HOP > CWiFi

if ws⇥ nAP ⇥ Cb + Crx(nAP ⇥ ws� 1) < Ctx

The only situation that verifies the above conditionsis to have a very high transmission cost compared tothe reception cost. In practice, transmission and recep-tion costs for low-power radios such as the CC2420radio are rather similar (Ctx

⇠= Cb⇠= 19 mA and

Crx⇠= 24 mA [22]). Hence, smart-HOP is expected

to be more efficient than the broadcast approach inWSNs and the conventional hand-off approaches inother wireless networks. The cost of the four hand-offapproaches is illustrated in Figure 3.

3.3 Test-bed setup for the preliminary experi-mentsThe aim of the preliminary experiments was to inves-tigate the feasibility of the smart-HOP mechanismin a controlled environment with limited dependen-cies on link dynamics. In this way, we deployed amodel-train in a large room (7 m⇥7 m) and the

4The beaconing (process of transmitting beacons) is definedseparately in order to be distinguished from the data transmission(C

tx

). However, the cost of transmitting a data packet and a beaconis assumed equal.

5Two more conditions of nAP

> 1 (existence of more than oneAP in the range of each MN) and ws > 1 (to apply a windowingprocess) are also respected.


2 4 6 8 100

100

200

300

400

500

600

Coom

unic

atio

n co

st (m

A)

Number of APs

smart−HOPBroadcastWiFiCellular

Fig. 3. The communication costs of broadcast inWSNs and the hand-off approaches in wireless net-works.

(a) (b)

Fig. 4. smart-HOP of the preliminary experiments forassessing and tuning the smart-HOP hand-off mech-anism (a) 4 APs and a MN, (b) MN passing by anAP [13].

locomotive followed a 3.5 m⇥3.5 m square layout (anextensive description on the preliminary experiments ispresented in [13]). The speed of the locomotive wasabout 1 m/s (similar to the average human walkingspeed). Figure 4(a) depicts the experimental scenarioand Figure 4(b) shows the locomotive passing by anAP. To prevent extreme deployment conditions suchas very high or very low density of APs, we guaran-teed a minimum overlap between neighboring APs.This was achieved by choosing a proper transmissionpower (-20 dBm) and locating the APs far enoughfrom each other.

The transmission period of the beacon and datapackets was 10 ms. This value is close to the maxi-mum rate possible, considering the processing, prop-agation and communication delays. The idea behindchoosing the maximum data rate was to evaluatesmart-HOP for scenarios with more demanding QoSrequirements. Four APs were located at the cornersof the railway, and up to six additional APs wererandomly placed, to assess the impact of AP’s density.

We ran four laps with the MN broadcasting packets,in each set of experiments. The experiments were runat different times of the day, during several days andwith a different number of people in the room. In allthese scenarios, the mobile node required a minimumof four hand-offs per lap. The time of the day andnumber of people in the room (1 to 4) did not seemto affect the number of hand-offs. We utilized aninterference-free channel to calibrate the parameters(channel 15, with a constant noise-floor of -94 dBm).

Performance metrics. In a mobile network, it iscrucial to maintain network connectivity as much aspossible by minimizing the inaccessibility periods and

TABLE 1Description of scenarios [13]

Scenarios Tl

HM m

A -95 dBm 1, 5 dBm 1, 2, 3C -85 dBm 1, 5 dBm 1, 2, 3B -90 dBm 1, 5 dBm 1, 2, 3D -80 dBm 1, 5 dBm 1, 2, 3

the frequency of hand-offs. In this line, we define thefollowing three metrics to evaluate the performanceof the smart-HOP algorithm.

1) Packet delivery ratio. It defines the ratio of packetssuccessfully delivered to the total number ofpackets sent.

2) Number of hand-offs. This metric helps identifyingthe existence of ping-pong effect. Multiple hand-offs in a single trip of a MN from one AP toanother AP means that the ping-pong effect hasoccurred.

3) Hand-off delay. It represents the network inacces-sibility time and is measured as the average timespent in the Discovery Phase (to find a better AP).Given that smart-HOP is a hard hand-off mech-anism, nodes cannot send packets during thistime; hence, this metric should be minimized.

3.4 Thresholds, hysteresis margin and AP stabil-ityThe first step in a hand-off scheme is to determinewhen should a node deem a link as weak and startlooking for another AP (represented as Tl in ourframework). In the sensor networks community, thede-facto way to classify links is to use the connected,transitional and disconnected regions.

An educated guess for the width of the hysteresismargin could be obtained from Figure 1 (based on the10 dBm width of the transitional region). However,while this value would guarantee that all links aboveTh are reliable, it would also increase the amountof beacons and time required to reach Th. In orderto evaluate this region extensively, we considereddifferent values for each hand-off parameter, as shownin Table 1. For example, if we consider scenario A witha 5 dBm margin and stability 2, it means that afterthe mobile node detects an AP above Th = �90 dBm,the node will send two 3-beacon bursts to observe ifthe link remains above Th. The hysteresis margin HMcaptures the sensitivity to the ping-pong effect, andthe number of bursts m reflects the stability of the APcandidate (recall that each burst in m contains threebeacons).

We conducted experiments for all the scenarios inTable 1. For each evaluation tuple < Tl, HM,m >,the mobile node performed four laps, leading to aminimum of 16 hand-offs. In each trip from one APto the next, an efficient scenario must perform onehand-off, which in turn leads to four hand-offs in onelap trip. The experiments provided some interestingresults. First, we show the results for the narrowmargin (1 dBm), and then the ones for the widermargin (5 dBm).


1 2 3 1 2 3 1 2 3 1 2 310

20

30

40

50

60

Num

ber o

f Han

doffs

Stability (m)

Scenario ATH=−95 dBm

DisconnectedRegion

TransitionalRegion

Scenario BTH=−90 dBm

Scenario DTH=−80 dBm

Scenario CTH=−85 dBm

1 2 3 1 2 3 1 2 3 1 2 30

2

4

6

8

Han

doff

Del

ay (s

)

Stability (m)





DisconnectedRegion

TransitionalRegion

1 2 3 1 2 3 1 2 3 1 2 350

60

70

80

90

100

Rel

ativ

e D

eliv

ery

Rat

io (%

)

Stability (m)

DisconnectedRegion

TransitionalRegion

Scenario DTH= −80 dBm

Scenario CTH= −85 dBm

Scenario BTH= −90 dBm

Scenario ATH= −95 dBm

(a) (b) (c)

Fig. 5. Results for narrow hysteresis margin (HM = 1dBm). (a) number of hand-offs, (b) mean hand-off delay,(c) relative delivery ratio. The horizontal lines represent the results for the best scenario: 32 for the number ofhand-offs and 96% for the relative delivery ratio [13].

3.5 ObservationsThe high variability of low-power links can causesevere ping-pong effect. Figure 5(a) depicts the totalnumber of hand-offs for the narrow margin case. Weobserved two important trends. First, all scenariosexhibit ping-pong effect. The minimum number ofhand-offs in this scenario is supposed to be 16 (onehand-off in each trip from one AP to another). How-ever, the figure indicates 32 to 48 hand-offs in 4-lapstrip. Due to the link variability, the transition betweenneighboring APs requires between 2 and 3 hand-offs.Second, a higher stability value m helps in alleviatingthe ping-pong effect. We observe that for all scenariosthe higher the stability, the lower the number of hand-offs.

Thresholds at the higher end of the transitionalregion lead to longer delays and lower deliveryrates. Figure 5(b) depicts the average hand-off delayfor various thresholds Tl. A threshold selected at thehigher end of the transitional region (-85 or -80 dBm,scenarios C and D) can lead to an order of magnitudemore delay than a threshold at the lower end (-90dBm, scenario B). This happens because mobile nodeswith higher thresholds spend more time looking foroverly reliable links (the Discovery Phase takes longer),and consequently less time transmitting data (lowerdelivery rate). Figure 5(c) depicts the relative deliv-ery rate and captures this trend. In order to have areference for the absolute delivery rate, we measuredseveral broadcast scenarios considering a high trans-mission rate and a 4-access point deployment. Wefound that the average delivery rate was 98.2%, witha standard deviation of 8.7. This implies that thereare limited segments with no coverage at all. Further-more, the overlap is minimal, which tests the agilityof the hand-off mechanism (as opposed to densedeployments, where very good links are abundant).Scenario A in Figure 5(c) is an exception, because theMN remains disconnected for some periods of time.As shown in Figure 4(a), no link goes below -95 dBm,hence, when this threshold is used, the Discovery Phasedoes not start because the link goes below Tl, butbecause disconnection time-outs occur.

The most efficient hand-offs seem to occur forthresholds at the lower end of the transitional region

and a hysteresis margin of 5 dBm. Figure 6 showsthat scenario B (-90 dBm) with stability 1 maximizesthe three metrics of interest. It leads to the lowestnumber of hand-offs, with the lowest average delayand highest delivery rate. It is important to highlightthe trends achieved by the wider hysteresis margin.First, the ping-pong effect is eliminated in all scenariosof Figure 6(a). Second, contrarily to the narrowerhysteresis margin, monitoring the stability of the newAP for longer periods (m = 2 or 3) does not provideany further gains, because the wider margin copeswith most of the link variability.

4 ANALYTICAL MODEL AND EVALUATIONThe performance of algorithms in low-power wirelessnetworks may greatly change depending on the net-work layout and environmental conditions. We con-ceived an analytical model for further evaluating thesmart-HOP algorithm. In this probabilistic analysis,we study the impact of two major channel parameters:

1) path-loss exponent (⌘). It measures the power ofradio frequency signals relative to distance.

2) standard deviation (�). It measures the standarddeviation in RSSI measurements due to log-normal shadowing.

The values of ⌘ and � change with the frequencyof operation and the clutter and disturbance in theenvironment. At this stage, we study the smart-HOPperformance in various environmental conditions toobserve the feasibility and efficiency of the algorithm.

In this section, first we describe the system modeland the probabilistic model for a hard hand-off pro-cess in WSNs. Then, we investigate the impact ofchannel parameters on the hand-off performance andcheck the analytical results through a simulation anal-ysis. After confirming the viability of the algorithm,we will move to more realistic experiments for bettertuning the relevant parameters (Section 5).

4.1 Probabilistic modelIt is important to consider a “probabilistic” or “analyt-ical” model that is faithful to the underlying physicalmodel while being amenable to analysis. We assumea scenario consisting of two APs (APa and APb) and


1 2 3 1 2 3 1 2 3 1 2 310

20

30

40

50

Num

ber o

f Han

doffs

Stability (m)

Disconnected Region

Transitional Region





1 2 3 1 2 3 1 2 3 1 2 30

1

2

3

4

5

6

7

8

Han

doff

Del

ay (s

)

Stability (m)

Transitional Region

Disconnected Region



Scenario BTH=−90 dBmScenario A

TH=−95 dBm

1 2 3 1 2 3 1 2 3 1 2 350

60

70

80

90

100

Rel

ativ

e D

eliv

ery

Rat

io (%

)

Stability (m)

DisconnectedRegion

TransitionalRegion

Scenario BTH= −90 dBm

Scenario CTH= −85 dBm

Scenario DTH= −80 dBm

Scenario ATH= −95 dBm

(a) (b) (c)

Fig. 6. Results for wide hysteresis margin (HM=5 dBm). (a) number of hand-offs, (b) mean hand-off delay, (c)relative delivery ratio. The horizontal lines represent the best results obtained for HM=1 dBm [13].

0 1 2 3 4 5−120

−100

−80

−60

−40

−20

Distance (m)

APa APb

THlow

THhighRSSI(dBm)

Threshold Level* RSSI of APaº RSSI of APb

Initialpositionof MN

Fig. 7. System model. It show the RSSI fluctuations ofAPa and APb in a 10 m distance. The threshold levelsare assumed -85 and -90 dBm. The hand-off happensin the middle of this trip (shadow area).

a MN. This assumption is enough without loss ofgenerality, as we are considering a hard hand-offprocess. In the probabilistic model, we ignore thewindow size and stability monitoring by consideringws=m=1, as adding these parameters into the analysisincreases the complexity of the equations and is notin the scope of our work.

The two main hand-off performance metrics arethe probability of ping-pong effect and the expectedhand-off delay. The system model and a general be-havior of smart-HOP are shown in Figure 7. The twoAPs are separated by distance d(m) while the MNmoves from the vicinity of AP a to the vicinity ofAP b along a straight line. In this model, the MNmoves with a constant speed of 1 m/s. The receivedsignal strength from AP a declines till it reaches thelower threshold level Tl, thus triggering the hand-off process. From this point onwards, the MN stopscommunication with APa and tracks the RSSI of theneighboring AP, AP b, (single radio eliminates theprobability of collecting RSSI readings from multipleneighbor APs at the same time). If the mobile nodeobserves a signal strength above a higher thresholdlevel, Th, the hand-off process is considered to befinished. The hand-off period is marked with a shad-owed vertical bar in Figure 7.

Link quality monitoring. There are different ways

of measuring the link quality metric. In this work,we consider the RSSI as the link quality level. Theprobabilities of being below the lower threshold leveland above the higher threshold level are defined byusing a Q-function. In this turn, the traveling path ofthe MN is divided into a number of slots. For the sakeof simplicity, we consider the same sampling rate forboth the Discovery and Data Transmission Phases. Theseprobabilities are expressed as follows.

P (Ra

(i) < Tl

) = Q(�T

l

+Ra

(i)�

)

P (Rb

(i) > Th

) = Q(Th

�Rb

(i)�

)

Where Q(.) is the complementary distributionfunction of the standard Gaussian, i.e., Q(x) =R1x

(1/p2⇡)e�t2/2dt, Ra(i) and Rb(i) indicates the

RSSI values from APa and APb at slot i, and � (indB) expresses the standard deviation.

Radio channel model. The received signal strengthis estimated by a log-normal shadowing path-loss.According to this model, R(i) (in dBm) (RSSI levelat a given slot i) from the transmitter is given by [23]:

R(i) = Pt

� PL(d0)� 10nlog10(i/d0)�X�

(1)

Where i corresponds to distance, Pt is the trans-mission power, PL(d0) is the measured path-loss atreference distance d0, n is the path-loss exponent, andX� = N(0,�) is a Normal variable (in dB). The termX� models the path-loss variation across all locationsat distance i from the source due to shadowing, a termthat encompasses signal strength variations due tothe characteristics of the environment (i.e., occlusions,reflections, etc.).

Smart-HOP probabilistic model. To evaluate theperformance metrics, we define the possibility of start-ing and ending a hand-off process at each slot. Asexplained earlier, in WSNs with hard hand-off, theMN communicates with a single AP at each time-slot.The MN is initially connected to APa —see Figure 7.When the MN is traveling from APa to APb, it tracksthe likelihood of the RSSI going below a thresholdlevel, Tl, at each sampling interval. By observinga low quality link, the hand-off process starts. Theprobability of starting a hand-off at slot s 2 [1, k) isdefined as follows (k indicates the total number of


slots).

P (S(s)) =

"s�1Y

i=1

P (Ra

(i) � Tl

)

#⇥ P (R

a

(s) < Tl

) (2)

The first part of the equation indicates the observa-tion of a number of slots (s�1) with good/acceptablelink quality level (above Tl). The second part denotesthe observation of the low link quality for the firsttime (below Tl). The following settings are used in allfuture evaluations across this section: � = 4 dB, ⌘ = 4,Pt = 0 dBm, d0 = 1 m, d = 5 m, PL(d0) = �55 dB,ws = m = 1, Tl = �90 dBm and Th = �85 dBm. Thenetwork related values are set according to the mostefficient scenario of the preliminary experiments.

By starting the hand-off process at slot/location s,the MN disconnects from the corresponding AP thatwas servicing the MN. At this moment, the MN startsassessing the other neighboring AP to choose the onewith higher threshold level (RSSI > Th). The hand-offfinishes when the MN observes a high link quality.Equation 6 formulates the probability of ending ahand-off at slot e considering the fact that the hand-off would have been started at slot s and the MN wasdisconnected from either APa or APb.

P (E(e) | S(s)) ="

e�1Y

i=s+1

(P (Ra

(i) < Th

)⇥ P (Rb

(i) < Th

))

#

⇥ [1� (P (Ra

(e) < Th

)⇥ P (Rb

(e) < Th

))] (3)

This equation assumes that the hand-off occurs atslot s. The ending moment at slot e 2 (s+ 1, k]happens at a later stage by comparing the RSSI levelof the APs to a higher threshold level, Th. Hence, inpractice it is a conditional probability and depends ona situation that has taken place previously at slot s.

The time span between the starting slot and theending slot is called hand-off delay. It is possible tocalculate the hand-off delay for all possible hand-offsstarting at any slot. By considering each point as astarting moment of a hand-off process, we character-ize the ending moments by the probabilities defined inEquation 6. The expected hand-off delay is computedby getting the weighted sum of all possible hand-offperiods. It is defined as the product of the time spentin each possible hand-off process started at slot s andended at slot e by the correspondent probabilities ofstarting a hand-off at slot s, P (S(s)), and ending itat slot e, P (E(e) | S(s)). For each hand-off startingat slot s, the hand-off would end at one of the slotsfrom s+1 to k. The sum of all these possible situationsdefines the expected delay for a hand-off started at aspecific slot s. The overall expected hand-off delay isdefined as follows.

Delay(s, e) =k�1X

s=1

kX

e=s+1

((e� s)⇥ P (E(e) | S(s))⇥ P (S(s))) (4)

In order to measure the ping-pong effect in smart-HOP, a new term is defined that is called probabilityof restarting a hand-off. This situation happens when

a MN performs hand-off at an improper moment, thusleading to an unnecessary hand-off. The restartingof a hand-off always occurs after successfully end-ing the first hand-off at slot r 2 (2, k]. This meansthat the probability of restarting is also a conditionalprobability that depends on ending a hand-off at anearlier stage. Since the MN may have been connectedto either APa or APb, the signal strength should beevaluated for both cases. The equation is defined asfollows.

P (R(r) | E(e)) =

"r�1Y

i=e+1

(1� P (Ra

(i) < Tl

)⇥ P (Rb

(i) < Tl

))

#

⇥ [P (Ra

(r) < Tl

)⇥ P (Rb

(r) < Tl

)] (5)

The second disconnection of the MN in this trip willbe ended at slot p, which is defined in the followingequation.

P (P (p) | R(r)) =

"p�1Y

i=r+1

(P (Ra

(i) < Th

)⇥ P (Rb

(i) < Th

))

#

⇥ [1� (P (Ra

(p) < Th

)⇥ P (Rb

(p) < Th

))] (6)

To find out the probability of ping-pong effect, thefull history of a MN since the first start of hand-off tothe end and restarting again are taken into account.Equation 7 illustrates the cases that lead to a ping-pong effect at slot p, as follows.

P (P (p)) =

" s�1Y

i=1

P (Ra

(i) < Tl

)

!⇥ P (R

a

(s) < Tl

)

#

⇥"

e�1Y

i=s+1

(P (Ra

(i) < Th

)⇥ P (Rb

(i) < Th

))

#

⇥ [1� P (Ra

(e) < Th

)⇥ P (Rb

(e) < Th

)]

⇥"

r�1Y

i=e+1

(1� P (Ra

(i) < Tl

)⇥ P (Rb

(i) < Tl

))

#

⇥ [P (Ra

(r) < Tl

)⇥ P (Rb

(r) < Tl

)]

⇥"

p�1Y

i=r+1

(P (Ra

(i) < Th

)⇥ P (Rb

(i) < Th

))

#

⇥ [1� P (Ra

(p) < Th

)⇥ P (Rb

(p) < Th

)] (7)

To find out the total probability of ping-pong effect,we define the following equation in a more abstractway. For each case of hand-off occurrence at slot s 2(0, k � 2], there is a chance to finish the hand-off atone of the upcoming slots e 2 (s, k � 1]. Similarly foreach e, as an ending slot, there is a chance of restartinganother hand-off at slot r 2 (e, k].

Total probability of ping � pong =k�3X

s

k�2X

e

k�1X

r

kX

p

P (S(s))⇥ P (E(e) | S(s))

⇥ P (R(r) | E(e))⇥ P (P (p) | R(r)) (8)

Performance metrics. To evaluate the performanceof the smart-HOP mechanism, two main metrics areconsidered, which are derived from the above equa-tions.


1) probability of ping-pong effect. It shows the proba-bility of reconnecting to APa after the first hand-off from APa to APb. This situation happens afterobserving a low quality link Rb(i) < Tl at APb

at slot i, where there was a high quality linkRb(i� 1) > Tl at slot i� 1.

2) expected hand-off delay. It indicates the expectedhand-off delay for each possible starting pointof hand-off from APa to APb.

In the following subsections, we study the impactof some parameters, which were either neglected ornot feasible to address due to the network limitationsin the preliminary experiments.

4.2 Impact of channel parametersAn increase in the path-loss exponent leads to longerhand-off delay and higher probability of ping-pongeffect. The path-loss exponent varies depending onthe environmental conditions. The path-loss parame-ter may be less dynamic in some applications with sta-tionary nodes and static environments or oppositelymay be highly variable in some other situations likemobile WSN applications [24]. Figure 8(a) illustratesthe variation of the RSSI at APa while the MN is mov-ing toward APb. The larger the path-loss exponent is,the higher the slope of the RSSI decrease will be. Insmart-HOP design, we aim at choosing a hand-offstarting level below the intersection of the receivedsignal power from APa and APb. This will reducethe chance of ping-pong effect. The ending level issupposed to be at this intersection or slightly higher.In practice, it is recommended to pick a higher levelfor ending point, to cancel out sudden fluctuations ofthe RSSI.

An increase in the shadowing variance enlargesthe transitional region, which in turn causes higherlink unreliability and ping-pong effect. This chan-nel parameter describes the received signal strengthfluctuation caused by flat fading. By increasing �, theprobability of entering the transitional region at closerdistances from the transmitter and leaving it at fartherdistances increases; this results in a larger transitionalregion —see Figure 8(b) [20].

By enlarging the channel parameters, the hand-offdelay increases. Figure 9 depicts a matrix, with thex-axis representing the path-loss exponent and the y-axis representing the shadow-fading. We observe anincreasing trend in the hand-off delay in each row andcolumn when increasing each channel parameter. Alarger path-loss exponent causes faster RSSI decrease.Hence, the MN enters the hand-off process at ear-lier stages. Finding a high link quality is postponedto later stages, which in turn increases the hand-off delay. Larger shadow-fading increases the RSSIstandard deviation, which expands the transitionalregion. Therefore, any disconnection from the pointof attachment requires a longer time assessing thewireless link to detect a high quality link in thetransitional region.

By enlarging the channel parameters, the proba-bility of ping-pong effect increases. Figure 10 showsthat increasing any channel parameter causes higher

0 2 4 6 8 10−110

−100

−90

−80

−70

−60

−50

−40

Distance (m)

RSS

I (dB

m)

η = 3η = 4η = 5T1 = −90 dBm

T2 = −85 dBm

(a)

(b)

Fig. 8. Impact of channel parameters on the RSSI withTl=-90 dBm and HM=5 dBm, (a) impact of path-lossexponent on RSSI, and (b) impact of shadow fading onthe RSSI [20].

Channel Parameters

η=1 η =2 η =3 η =4 η =5 η =6

σ=1 8.2e-099 1.4e-045 2.4e-014 0.0216 0.0387 0.0469 σ=2 3.1e-027 9.1e-014 2.5e-005 0.0324 0.0408 0.0470 σ=3 2.4e-013 2.8e-007 0.0022 0.0371 0.0422 0.0477 σ=4 2.6e-008 7.5e-005 0.0109 0.0398 0.0433 0.0496 σ=5 6.9e-006 0.0012 0.0215 0.0415 0.0443 0.0519 σ=6 1.6e-004 .0054 0.0297 0.0428 0.0453 0.0537

Fig. 9. Impact of channel parameters on the overallexpected hand-off delay in seconds (sampling rate ofevery 50 ms).

link variability, unreliability and instability. This isthe main reason for noticing higher probability ofping-pong effect when increasing either the path-lossexponent or shadow fading parameters.

Studying the channel parameters (�,⌘) reveals thehigh dependency of the hand-off process on envi-ronmental changes. However, these values do notfluctuate significantly in indoor environments [20].In an efficient hand-off, the MN should perform theprocess within at most a single sample (50 ms inthis example). Figure 9 shows an acceptable hand-off delay for most cases except with ⌘ = 6 and� > 4. This condition rarely happens in outdoorenvironments [20]. Thus, we get to the conclusionthat smart-HOP is suitable for all environments,although for outdoor environments a user shouldperform a radio survey6 to obtain a better insight.

6It is difficult to predict the values of channel parametersduring an experiment. A radio survey is a process that determinesthe channel values before performing an experiment.


Channel Parameters

η =1 η =2 η =3 η =4 η =5 η =6

σ=1 0 0 1.8e-229 6.4e-144 5.4e-079 2.1e-034 σ=2 5.7e-120 3.7e-087 5.3e-060 1.7e-038 6.6e-022 2.2e-010 σ=3 1.0e-054 4.2e-040 5.7e-028 2.7e-018 1.0e-010 2.4e-005 σ=4 1.6e-031 2.7e-023 1.7e-016 6.8e-011 1.6e-006 0.0018 σ=5 1.4e-020 2.6e-015 5.5e-011 2.7e-007 1.7e-004 0.0128 σ=6 1.7e-014 7.2e-011 7.3e-008 2.8e-005 0.0022 0.0345

Fig. 10. Impact of channel parameters on the proba-bility of ping-pong effect (sampling rate of 20 Hz).

1 2 3 4 50.1

0.2

0.3

0.4

Window size

Han

d−of

f del

ay (s

)

Handoff delay

0.98

1

1.02

Ave

rage

num

ber o

f han

d−of

fs

Avg. no. of hand−offs

1 2 3 4 50.2

0.4

0.6

0.8

1

1.2

Stability monitoring

Han

d−of

f del

ay (s

)

Handoff delay

0.98

1

1.02

Ave

rage

num

ber o

f han

d−of

fs

Avg. no. of hand−offs

(a) (b)

Fig. 11. Simulation analysis; (a) impact of windowsize in the Data Transmission Phase, and (b) impactof stability monitoring.

4.3 Simulation modelWe performed a simulation study with MATLAB toverify the correctness of the probabilistic model [25].In this model, we generated random values of RSSI atvarious distances from the serving AP and the neigh-boring AP with Equation 1. The mobile node startedand ended the Discovery Phase by reading the RSSIvalues at each sampling slot. Studying the impact ofnetwork and channel parameters (Tl, HM , m, ⌘ and�), we observed similar results to the ones from theprobabilistic model. In the simulation model, we areable to consider higher values of ws and m. Theseparameters were ignored in the probabilistic model,for simplicity. We assume that the MN is initiallyconnected to APa. By considering a sliding windowws and low threshold level Tl for starting a hand-off,the MN decides for the hand-off starting slot. Thenby having the RSSI value of APb and considering thestability parameter m, the MN decides for ending thehand-off. This process repeats for 10,000 trips and theresults are averaged at the end of the simulation.

Impact of window size. This parameter is used inboth the Discovery Phase and the Data TransmissionPhase. In the preliminary experiments, we simplyassumed ws=3 for both phases. In this simulation,we study the impact of window size in each phaseseparately.

Impact of window size in the Data Transmission Phase.By setting ws=3 for the Discovery Phase and varyingit from 1 to 5 in the Data Transmission Phase, we get adecreasing trend of hand-off delay for the first 3 casesand then it remains unchanged —see Figure 11(a).This happens to the number of hand-offs as well.This means that a small window size value during thenormal data communication of the MN, reduces thehand-off delay and the number of unnecessary hand-offs.

Impact of window size in the Discovery Phase. Increas-ing the number of beacons for assessing the neighbor-ing APs requires more time, which is proportional tothe sampling frequency. This case is somehow similarto the stability parameter that increases the period oflink assessment. It is apparent that considering a fewsamples can compensate the fluctuations of randomRSSI values. Hence, it is not logical to assume a largewindow size value for the Discovery Phase due to itsnegative impact on the hand-off delay. In the currentmodel with 2 APs, the result is similar to the casewhen changing the stability parameter, thus it is notshown here. In case of higher density scenarios theresults are different, but still the trend is equal.

Impact of stability monitoring. Increasing the sta-bility monitoring reduces the link variability. Each unitof stability monitoring adds a new Discovery Phase,which is composed of a set of beacons and replypackets. The results in Figure 11(b) indicate that thehand-off delay has an increasing trend with a highslope, which is more steep than the case of increasingwindow size during the Discovery Phase. Consideringthe ping-pong effect, there is an improvement with asmall stability parameter. In practice, we can substi-tute the stability parameter with the window size of theDiscovery Phase. By this action, we can (i) reduce thelink variability to eliminate the ping-pong effect and(ii) compensate the RSSI fluctuations to take accuratehand-off decisions. More insight into the simulationis provided in [14].

5 EXPERIMENTAL EVALUATIONThe preliminary experiments [13] revealed the bestthresholds for a hand-off process in a controlledenvironment. The probabilistic analysis proved that,in theory, smart-HOP is able to perform efficienthand-offs. At this stage, we set further experimentsto enable a deeper analysis and fine-tuning of thealgorithm in a more realistic environment.

We deployed a maximum of 6 APs with a minimumpower level of -25 dBm in a 80 m2 room. The APswere attached to walls at 1.5 m height from theground (to guarantee a better connectivity). Figure 12illustrates the position of each AP, furniture, walls andwindows. A person was holding the mobile node andthe logging PC7.

5.1 Test-bed setupAs we mentioned earlier, the two parameters of lowthreshold level and hysteresis margin are very impor-tant. Instead of starting the experiment with all the6 APs, we first confine the scenario to 2 APs (AP1

and AP2 in Figure 12) and attached the MN to theperson’s shoulder, which faces the anchors in eachtrip from AP1 to AP2. On the way back, the bodyeliminates the Line-of-Sight (LoS) communication. We

7At the beginning, we connected all APs to one laptop withpassive USB cables and USB2.0 hubs. Then we observed some dataloss during data transfer through the UART port. Adding morePCs did not solve the problem completely. Hence, we managed toget the data log from the MN with the cost of a person carrying alaptop during the experiment.


Kitchen

Furniture

Furniture Furniture

AP6

AP1 AP2

AP5 AP4

AP3

Fig. 12. The APs’ deployment in a large room.

refer to this set of tests as baseline experiments. LessAPs guarantees that there is no overlapping betweenneighboring APs. The person walks 4 times betweenthese APs with a normal human walking speed (about1 m/s).

The experimental area is a lab with at most10 people sitting and three to five people movingrandomly in all locations. The tests were performedon channel 15 of the CC2420 radio, which maybeaffected by different sources of interference in the 2.4GHz band (such as WiFi, Bluetooth and microwavedevices). In order to obtain a better understandingof the frequency activities, we measured the 2.4 GHzspectrum usage; a WiFi-Spy spectrum analyzer veri-fied that there was very low interference from otherRF sources.

We evaluate smart-HOP in this realistic environ-ment in two steps; (i) baseline experiments using 2 APs,to further analyze the lower threshold level and thehysteresis margin and (ii) extended experiments using6 APs, to study the impact of stability monitoringand window size. In all tests, we employ SNR-basedsmart-HOP as it encompasses the interference in theenvironment.

5.2 Evaluation - baseline experimentsIn the preliminary experiments (see Section 3), we con-sidered four groups of lower threshold level (-95, -90,-85 and -80 dBm) with 2 values of hysteresis margin 1and 5 dBm. The results indicated that -95 dBm is nota choice as the MN enters in the disconnected region.Now, we consider a wider range of lower thresholdlevels [-76, -90 dBm] increasing in 2 dBm steps, andhigher threshold levels in the range of [-75,-89 dBm],which in turn generate hysteresis margin ranging from1 to 15 dBm. All the 8 cases of lower threshold levelswith variations of HM lead to 36 combinations. Wecompare all these situations in terms of number ofhand-offs, hand-off delay and packet delivery ratio,walking four times between APs (APa and APb). Themain goal at this stage is to pick situations that aremore likely to be efficient and then reassess themwith 6 APs, for further comparison. We observe thefollowing facts.

1) Selecting the lower threshold level from thelower end of the transitional region with a widerhysteresis margin eliminates the ping-pong effect.

2) Either very wide hysteresis margin or narrowmargin with large threshold level causes hugehand-off delay. A wide hysteresis margin obligesthe MN to stay at the Discovery Phase for longer

periods of time. A narrow hysteresis margin withlarge value of lower threshold level causes anexcessive number of hand-offs that eventuallyenlarges the hand-off delay.

3) A lower hand-off delay causes higher packetdelivery ratio. The evaluations revealed that(i) increasing the hysteresis margin in all casesreduces the link variability and increases thepacket delivery ratio, and (ii) higher values ofpacket delivery are achieved in situations withlower hand-off delay. The more efficient scenar-ios are noticed with HM between 3 and 7 dBmand Tl in the range of -86 to -90 dBm.

The most efficient scenarios in the baseline exper-iments are more elaborated in the extended experi-ments. The baseline experiments reveal that with thesmallest Tl, -90 dBm, and HM=5–7 dBm, the hand-off delay is minimum, while obtaining a maximumdelivery of packets. An educated solution is to keepthe MN connected to the current link as much aspossible, similarly to the preliminary experiments. Thus,we keep the same threshold level for starting thehand-off (Tl=-90 dBm) and compare the results ofvarious HM values (3 to 8 dBm).

5.3 Evaluation - extended experimentsTo find the best setting for the hysteresis margin, weincrease the number of APs to six. Adding moreAPs creates a more realistic environment in whichthe mobile node experiences links overlapping. Foreach set of experiments, according to the selectedhysteresis margin, the person walks in the room whilethe mobile node sends data periodically (every 100ms). The person starts walking from AP1, along thedashed line shown in Figure 12. In some parts of theway, there are obstacles that prevent Line of Sight(LoS) communication. Moreover, random movementof people creates more dynamics in the environment.For each set of experiments, the person walked forabout 15 minutes (transmitting 10,000 packets), whichis about 15 full laps (dashed line circuit). At the end,we computed the average hand-off delay and thepacket delivery ratio.

Increasing the hysteresis margin enlarges the hand-off delay as it forces the MN to attach to a higherlink quality AP. Figure 13(a) shows that the hand-off delay is minimum with 3-5 dBm hysteresis marginand then it records a gradual increase. The reason isthat by enlarging the margin, the chance of stayingin the Discovery Phase for more than “one period”is higher. The one period stands for the case whereafter sending a burst of beacons and receiving replypackets, the MN is able to observe a good link to makethe hand-off. The packet delivery ratio in Figure 13(b)illustrates a higher packet delivery ratio with HM=5and 6 dBm. The packet delivery decreases with HMsince there are unnecessary hand-offs. The higher HMcauses fewer packets delivered as the MN stays in theDiscovery Phase longer.

The hysteresis margin should be tuned to achievean optimal trade-off between the delivery rate anddelay. By choosing the lower end of the transitional


−87 −86 −85 −84 −83 −820

50

100

150

200

250

Han

d−of

f del

ay (m

s)

THhigh (dBm)−87 −86 −85 −84 −83 −8280

85

90

95

100

Pack

et d

eliv

ery

ratio

(%)

THhigh (dBm)

(a) (b)

Fig. 13. More extensive experimental evaluation with 6APs by considering Tl=-90 dBm, (a) average hand-offdelay. The horizontal line shows the optimum hand-offdelay computed by the analytical model. (b) absolutepacket delivery ratio.

region (-90 dBm) as the threshold level, the hysteresismargin of 5 dBm is the best choice. The stabilitymonitoring and the window size parameters are thetwo other important parameters, which are studiedin these experiments.

Increasing the stability monitoring increases thehand-off delay. Increasing the stability monitoring hasa direct impact on the hand-off delay —see Fig-ure 14(a). Adding one unit to the stability requiresobserving a high quality link for one more slidingwindow. It is interesting to notice that this raisedoes not have a good impact on the packet deliverysince we are shrinking the Data Transmission Phase.Considering the fact that a slight change in stabilityincreases the hand-off delay significantly, it is wise totune other related parameters with less impact. Thus,we opt for choosing the minimum stability value forthe experiment and play with other parameters.

The window size parameter compensates the dy-namics of the link. It should neither be too smallnor too large as depicted in Figure 14(b). This pa-rameter compensates the link variability and suddenRSSI changes. In a controlled environment, ws=3 wasselected, according to the suggestions from relatedwork. However, in a realistic scenario, there are moresources of disturbance: (i) there is a natural variationin human gait. When a movement experiment isrepeated, a person carrying a node may move a bitfaster or slower than before, or may deviate slightlyfrom the previous path. Hence, a node may detectdifferent signal strength at the same position [26].(ii) The human body partly absorbs electromagneticradiation, and the amount of absorbed energy de-pends –among other things– on the person’s physiqueand pose, and the radio frequency [27]. The resultsindicate that, in a real environment, a slightly higherwindow size, ws=4, increases the accuracy in termsof hand-off decision on exact moments. But enlargingmore than this value does not improve the perfor-mance since it provides coarse grain information ofthe link. Considering wider window sizes reduces theresponsiveness of the hand-off process.

1 2 3 4 5100

200

300

400

500

600

700

Stability monitoring

Han

d−of

f del

ay (m

s)

Handoff delay

82

84

86

88

90

92

94

Pack

et d

eliv

ery

ratio

(%)

Packet delivery ratio

3 4 5 6 7120

125

130

135

140

145

Window size

Han

d−of

f del

ay (m

s)

Handoff delay

88

90

92

94

96

98

Pack

et d

eliv

ery

ratio

(%)

Packet delivery ratio

(a) (b)

Fig. 14. (a) Impact of stability monitoring, (b) impact ofwindow size.

6 BACKGROUND AND RELATED WORK

Hand-off mechanisms have been widely studiedin cellular [28]–[30] and wireless local area net-works [31]–[33], but did not receive the same levelof attention in low-power wireless networks. Thereare two major strategies for the hand-off process: softhand-off (network layer) and hard hand-off (MACsub-layer). The soft hand-off requires lots of packet ex-changes, thus impacting the sustainability of energy-constrained nodes. In any case, we address some ofthe latest algorithms using soft hand-off in WSNs [34],[35] as follows.

In [34], the authors focus on the hand-off in net-works with mobile sensors and gateways. The mobilenode is supposed to be in the range of multiplegateways that are periodically broadcasting routeradvertisement (RA) packets. Advertising their pres-ence, enables the mobile node to decide for the bestgateway. The connectivity of the network relies onthe RA packets frequency; high frequency leads tonetwork congestion while low frequency leads to lowresponsiveness of the network (thus to longer networkinaccessibility).

A soft hand-off within 6LowPAN is proposedin [35]. The paper claims zero hand-off time andzero packet losses. The process is similar to [34].However, it takes advantage of using two additionalcontrol messages, namely Join and Join Ack that aresent/received when the MN is still attached to theserving AP. This algorithm requires a huge amount ofcontrol message exchanges, increasing the probabilityof network congestion. Moreover, the zero hand-offdelay was observed in a low sampling rate scenario,which may not be the case in many applications.

A more reasonable approach for low-power wire-less networks (hard hand-off) is based on data linklayer solutions [9], [13]. The respective authors claimthat their approaches are adequate for passive deci-sion with non-real-time support in [9] or for activedecision with real-time support in [13].

In [9], the authors describe a wireless clinical mon-itoring system collecting vital signs from patients. Inthis study, the mobile node connects to a fixed AP bylistening to beacons periodically broadcast by all APs.The node connects to the AP with the highest RSSI.The scheme is simple and reliable for low traffic datarates. However, there is a high utilization of band-width due to periodic broadcasts (similar to soft hand-offs) and hand-offs are passively performed whenever


the mobile node cannot deliver data packets.We proposed a more reliable and faster hard hand-

off approach for WSNs [13]. In this algorithm, hand-off is initiated at the mobile node, opposed to otheralgorithms in the literature. The MN keeps track of thelink quality level during the Data Transmission Phase.A timer is responsible to detect the unreachability ofthe serving AP. Hence, the MN is able to detect thelink degradation and unreachability of the serving APwithin a short time. Then, the MN spends a timewindow assessing the neighboring APs to change tothe best one.

The link quality is one of the parameters that sig-nificantly affects the hand-off performance. Differentlink quality estimators have been proposed for sensornetworks. They apply different criteria to estimate thelink status, such as RSSI, SNR, LQI or link asymme-try [36], [37]. In our case, due to the dynamics im-posed by mobility, we use a simple and fast samplingof RSSI and SNR, which have been shown to providereliable metrics [15].

7 CONCLUDING REMARKSTo the best of our knowledge, we are the first tosystematically and empirically evaluate the hand-offin low-power wireless networks. We believe that thisis important, because a panoply of WSN applica-tions may require mobile nodes to report informationreliably and in real-time, such as in clinical healthmonitoring and industrial automation. We proposed areliable hand-off procedure that we dubbed as smart-HOP [13]. This hand-off scheme enables the MN todeliver data through the neighboring AP that offersthe best link quality. In this line, some key parametershave a more significant impact on the performanceof the hand-off. Experimental results in a controlledenvironment revealed the best threshold level (Tl=-90 dBm), hysteresis margin (HM=5 dBm) and stabilitymonitoring (m=1) values, achieving hand-off delays offew tens of milliseconds and relative packet deliveryratios of around 98%.

It is well known that the performance of a low-power wireless system is very prone to environmentalconditions. Thus, we conceived a probabilistic modelto investigate the impact of the most relevant channelparameters on the hand-off process. We showed thatthe environmental changes have a direct impact on thesmart-HOP performance, but it ends up performingwell in various channel conditions. To have a betterknowledge of the smart-HOP performance, it is rec-ommended to perform a radio survey (to determinethe path-loss exponent and shadowing standard de-viation values) before the experiment. A simulationmodel was also designed to verify the probabilisticmodel. We studied the impact of network and channelparameters, confirming the correctness of the proba-bilistic analysis. The impact of window size and thestability monitoring parameters were also investigated.It was revealed that the stability monitoring has muchmore strength that the hysteresis margin in what con-cerns the hand-off delay.

We performed an extensive set of experiments ina more realistic environment. These were performed

in a large room with more people around, while theMN was attached to the shoulder of a person and(two to six) access points were attached to walls. Awider range of parameters was considered for theperformance analysis (Tl from -90 to -76 dB, HMfrom 1 to 15 dBm, m from 1 to 5 and ws from 3to 7). We obtained similar parameter settings as inthe preliminary experiments (in a controlled environ-ment), which confirmed the stability of the smart-HOP mechanism in various environmental conditionsand for several network scenarios.

The smart-HOP design shows some advantages butalso limitations. It enables fast and reliable mobilitysupport in low-power networks. It requires a numberof stationary APs that are deployed in such a way toprovide minimum overlap. In a dense deployment,the MN is very likely in the connected region ofone AP, (which rarely happens in WSN applications).smart-HOP is inefficient for dense deployments as it istuned based on the assumption of existing transitionalregions. Moreover, the single radio characteristic lim-its the number of MNs that can be serviced at eachAP. In cellular networks, each base station supportshundreds of MNs. The 802.15.4 radio allows onecommunication at each instance of time that limits thenumber of MNs.

This paper described the design and implementa-tion of smart-HOP in a “protocol-agnostic way”withone MN and a number of APs. A future directionof this work is to support the smart-HOP withinstandard protocols and commercial off-the-shelf tech-nologies (e.g. 6LoWPAN).

ACKNOWLEDGMENTSThe authors thank Patrick Meumeu Yomsi for hisgreat help in revising the analytical model. Thiswork was partially supported by National Fundsthrough FCT (Portuguese Foundation for Science andTechnology) and by ERDF (European Regional De-velopment Fund) through COMPETE (OperationalProgramme ’Thematic Factors of Competitiveness’),within projects FCOMP-01-0124-FEDER-037281 (CIS-TER), FCOMP-01-0124-FEDER-014922 (MASQOTS);also by FCT and the EU ARTEMIS JU funding withinARROWHEAD project, ref. ARTEMIS/0001/2012, JUgrant 332987.

REFERENCES[1] (2013) Final report from the nsf workshop on future

directions in wireless networking. [Online]. Available: http://ecedha.org/docs/nsf-nets/final-report.pdf?sfvrsn=0

[2] (2013) European commission, “factories of the future 2020:multi-annual roadmap for the contractual ppp underhorizon 2020”, research roadmap produced by the europeanfactories of the future research association (effra). [Online].Available: http://www.effra.eu/attachments/article/129/Factories%20of%20the%20Future%202020%20Roadmap.pdf

[3] P. J. Marron, D. Minder, and S. Karnouskos, The EmergingDomain of Cooperating Objects. Springer, 2012.

[4] (2012) Ginseng: performance control in wireless sensornetworks. [Online]. Available: http://www.ucc.ie/en/misl/research/previous/ginseng/

[5] (2009) Fasys: Absolutely safe and healthy factory. [Online].Available: http://www.fasys.es/en/proyecto.php

[6] (2010) flexware : Flexible wireless automation in real-timeenvironments. [Online]. Available: http://www.flexware.at/


[7] J. Martinez-de Dios, A. Jimenez-Gonzalez, A. San Bernabe, andA. Ollero, “Conet integrated testbed architecture,” in A RemoteIntegrated Testbed for Cooperating Objects, ser. SpringerBriefs inElectrical and Computer Engineering. Springer InternationalPublishing, 2014, pp. 23–39. This work has been carried outwithin the Cooperating Objects European Network of Excel-lence http://www.cooperating–objects.eu/.

[8] R. Silva, J. Silva, and F. Boavida, “A proposal for proxy-basedmobility in wsns,” Computer Communications, 35(10), 2012.

[9] O. Chipara, C. Lu, T. C. Bailey, and G.-C. Roman, “Reliableclinical monitoring using wireless sensor networks: Experi-ences in a step-down hospital unit,” in ACM SenSys, 2010.

[10] K. Lorincz, D. Malan, T. Fulford-Jones, A. Nawoj, A. Clavel,V. Shnayder, G. Mainland, M. Welsh, and S. Moulton, “Sensornetworks for emergency response: challenges and opportuni-ties,” IEEE Trans. Pervasive Computing, 3(4), pp. 16–23, Oct 2004.

[11] X. Li, R. Falcon, A. Nayak, and I. Stojmenovic, “Servicingwireless sensor networks by mobile robots,” IEEE Comm. Mag.,50(7), 2012.

[12] Y. Pei and M. W. Mutka, “Stars: Static relays for multi-robotreal-time search and monitoring,” in IEEE DCOSS, 2011.

[13] H. Fotouhi, M. Zuniga, M. Alves, A. Koubaa, and P. Marron,“Smart-hop: A reliable handoff mechanism for mobile wirelesssensor networks,” in EWSN, 2012.

[14] H. Fotouhi, M. Alves, M. Zuniga, and A. Koubaa,“Reliable and fast hand-offs in low-power wirelessnetworks (extended paper),” Tech. Rep., 2014. [Online].Available: http://www.cister.isep.ipp.pt/docs/reliable+and+fast+hand%252Doffs+in+low%252Dpower+wireless+networks++%2528extended+paper%2529/825/view.pdf

[15] M. Zuniga, I. Irzynska, J. Hauer, T. Voigt, C. Boano, andK. Roemer, “Link quality ranking: Getting the best out ofunreliable links,” in IEEE DCOSS, 2011.

[16] N. Baccour, A. Koubaa, L. Mottola, M. Zuniga, H. Youssef,C. Boano, and M. Alves, “Radio link quality estimation inwireless sensor networks: a survey,” ACM Trans. Sensor Net-works, 8(4), pp. 34:1–34:33, 2012.

[17] A. Woo, T. Tong, and D. Culler, “Taming the underlyingchallenges of reliable multihop routing in sensor networks,”in ACM SenSys, 2003.

[18] L. Mottola, G. P. Picco, M. Ceriotti, S. Guna, and A. L.Murphy, “Not all wireless sensor networks are created equal:A comparative study on tunnels,” ACM Trans. Sensor Networks,7(2), pp. 15:1–15:33, 2010.

[19] B. Kusy, C. Richter, W. Hu, M. Afanasyev, R. Jurdak, M. Brunig,D. Abbott, C. Huynh, and D. Ostry, “Radio diversity forreliable communication in wsns,” in ACM/IEEE IPSN, 2011.

[20] M. Zuniga and B. Krishnamachari, “Analyzing the transitionalregion in low power wireless links,” in IEEE SECON, 2004.

[21] M. Zuniga and B. Krishnamachari, “An analysis of unreliabil-ity and asymmetry in low-power wireless links,” ACM Trans.Sensor Networks, 3(2), pp. 1–34, 2007.

[22] J. Polastre, R. Szewczyk, and D. Culler, “Telos: enabling ultra-low power wireless research,” in ACM/IEEE IPSN, 2005.

[23] T. Rappaport, Wireless communications: principles and practice.IEEE press, 1996.

[24] N. Alam, A. Balaie, and A. Dempster, “Dynamic path lossexponent and distance estimation in a vehicular network usingdoppler effect and received signal strength,” in IEEE VTC,2010.

[25] M. Thamri, “Evaluation of smart-hop: a handoff approachfor mobile wireless sensor networks,” Master’s thesis, HigherSchool of Communication of Tunis, Sup’Com, 2012.

[26] R. Fish, M. Flickinger, and J. Lepreau, “Mobile emulab: Arobotic wireless and sensor network testbed,” in IEEE INFO-COM, 2006.

[27] O. Rensfelt, F. Hermans, P. Gunningberg, L. Larzon, andE. Bjornemo, “Repeatable experiments with mobile nodes in arelocatable wsn testbed,” The Computer Journal, 54(12), 2011.

[28] S. Cho, E. Jang, and J. Cioffi, “Handover in multihop cellularnetworks,” IEEE Comm. Mag., 47(7), 2009.

[29] T. Salih and K. Fidanboylu, “Modeling and analysis of queu-ing handoff calls in single and two-tier cellular networks,”Computer Communications, 29(17), 2006.

[30] B. Madan, S. Dharmaraja, and K. Trivedi, “Combined guardchannel and mobile-assisted handoff for cellular networks,”IEEE Trans. Vehicular Tech., 57(1), 2008.

[31] I. Ramani and S. Savage, “Syncscan: practical fast handoff for802.11 infrastructure networks,” in IEEE INFOCOM, 2005, pp.675–684 vol. 1.

[32] A. Mishra, M. Shin, and W. Arbaugh, “An empirical analysis ofthe ieee 802.11 mac layer handoff process,” ACM SIGCOMM,2003.

[33] M. Shin, A. Mishra, and W. A. Arbaugh, “Improving thelatency of 802.11 hand-offs using neighbor graphs,” in ACMMobiSys, 2004.

[34] J. Petajajarvi and H. Karvonen, “Soft handover method formobile wireless sensor networks based on 6lowpan,” in IEEEDCOSS, 2011.

[35] Z. Zinonos and V. Vassiliou, “S-ginmob: Soft-handoff solutionfor mobile users in industrial environments,” in IEEE DCOSS,2011.

[36] D. De Couto, D. Aguayo, J. Bicket, and R. Morris, “A high-throughput path metric for multi-hop wireless routing,” inACM MobiCom, 2003.

[37] N. Baccour, A. Koubaa, H. Youssef, M. Ben Jamaa,D. do Rosario, M. Alves, and L. Becker, “F-lqe: A fuzzy linkquality estimator for wireless sensor networks,” in EWSN,2010.

Hossein Fotouhi is a Ph.D. student atCISTER/INESC-TEC Research Unit andUniversity of Porto, working in the areas ofsensor networks, mobile computing and In-ternet of Things. In 2009, he completed Mas-ter of Science at University Putra Malaysiaon Communication Network Engineering. In2004, he obtained Bachelor of Science atUniversity of Guilan on Electrical ElectronicsEngineering. He worked 3 years afterward inindustry as a network engineer.

Mario Alves has a PhD (2003) in ECE atthe University of Porto, Portugal. He is aProfessor in ECE at Politecnico do Porto(ISEP/IPP) and a Research Associate at theCISTER/INESC-TEC Research Unit, a top-ranked Portuguese research centre. His cur-rent research interests are mainly devoted toquality-of-service (QoS) in low-power wire-less networks, particularly concerning time-liness and reliability issues.

Marco Zuniga Zamalloa is an AssistantProfessor in the Department of ComputerScience at Delft University of Technology,the Netherlands. He obtained his PhD andMSc in Electrical Engineering from the Uni-versity of Southern California, in 2002 and2006, respectively; and his BSc in Electron-ics Engineering from the Pontificia Universi-dad Catolica del Peru. His research interestare in the areas of wireless networks, perva-sive computing and cyber physical systems.

Anis Koubaa is currently an Associate Pro-fessor in Computer Science at Prince SultanUniversity and Senior Research Associateat the CISTER/INESC-TEC Research Unit.He has a PhD (2004) in Computer Sciencefrom INPL Lorraine, France. He edited andauthored more than 4 books, 90 journal andconference papers. He received the award ofthe Best Research Work competition from Al-Imam University in 2010. His research inter-ests are mobile robots and sensor networks.

Date post:	09-Mar-2023
Category:	Documents
Upload:	mdh
View:	1 times
Download:	0 times

Reliable and fast hand-offs in low-power wireless networks

Documents